Coding Agents From Tencent Nearly Double Success to 72.7%

Tencent’s new research paper, flagged Saturday by AI researcher Elvis Saravia, describes a way for terminal and coding agents to refine their own training without constant human tuning.

Key Takeaways

Tencent’s paper describes agents generating their own training signals by grading coding-task attempts and using those judgments in later training rounds
Terminal agents can write and execute code, run tests and fix errors in a loop within command-line or development environments
Autonomous machine learning engineering systems increased completion rates on ML engineering benchmarks from 37.9% to 72.7% in separate research
Researchers are assessing whether recursive self-improvement generalizes beyond coding benchmarks and can avoid optimizing for the wrong signal

Saravia flagged the paper to his followers on Saturday and called it a “banger paper” worth bookmarking for anyone tracking how terminal and coding agents are trained with reinforcement learning, a training method that rewards a model for successful actions rather than showing it labeled examples.

Terminal agents are AI systems built to operate a command line or development environment directly, writing and executing code, running tests and fixing their own errors in a loop rather than just suggesting snippets for a human to copy.

Recursive self-improvement, often shortened to RSI, describes a system that uses its own outputs to upgrade the process that produced them, rather than waiting for engineers to manually redesign the training pipeline.

What Recursive Self-Improvement Actually Changes

The Tencent paper describes a recursive self-improvement approach that lets an agent generate its own training signal, effectively grading its own attempts at a coding task and feeding that judgment back into the next training round.

For Tencent’s method, the practical effect, if it holds up outside a lab benchmark, is a shorter gap between a model failing at a task and a model that no longer fails at it, with less need for a human to sit in that loop.

Separate work on autonomous machine learning engineering, published by researchers studying agentic systems that develop and improve AI pipelines without supervision, found such systems nearly doubling their completion rate on ML engineering benchmarks, climbing from 37.9% to 72.7%.

That kind of jump is the practical payoff recursive self-improvement research is chasing, fewer tasks where a human has to step in and finish what the agent started.

Also Read: Kim Jong Un’s North Korea Claims First AI-Guided Torpedo, No Proof Offered

A Research Line Moving From Labs To Deployment

Interest in self-improving agents accelerated through 2026 as labs including OpenAI and Google DeepMind pushed coding agents further into autonomous territory. Benchmark papers arrived almost weekly comparing how models handle multi-step software engineering tasks without supervision.

Tencent, China’s largest gaming and social media company and a major AI research funder, has increasingly published alongside Western labs on agentic training methods rather than consumer-facing model releases, positioning its research arm as a contributor to the same open literature US labs draw from.

Why Researchers Are Watching The Self-Improvement Curve

The open question is whether recursive self-improvement methods generalize beyond the narrow coding benchmarks they are tested on, or whether they need careful guardrails to stop agents from optimizing for the wrong signal.

Anthropic and OpenAI have both published safety material this month on monitoring agent behavior, precisely because autonomous training loops are harder to audit than human-supervised fine-tuning.

Watch for follow-up replications or critiques of the Tencent method over the coming weeks. Those will determine whether the technique becomes a standard part of how coding agents are trained, or stays a lab curiosity confined to benchmark papers.